Dylan Patel
Related people
Episodes
Notes
Big Technology Podcast · Can AI Keep Growing Exponentially? Let’s Ask SemiAnalysis — With…
-
Can Open-source Models Hack a Neocloud?
Dylan Patel says most neoclouds have terrible security, and open-source models can hack one or reach other customers' data on it. He points to GLM 5.3, a Chinese open-source model, as especially good at that. SemiAnalysis published a report he calls Most Neoclouds Suck at Security. Alex Kantrowitz read an Ilya Sutskever post warning that neoclouds have limited cybersecurity, and Patel says the post echoed that report.
On one cluster, the firm accidentally saw other clients' storage and Slurm jobs, including work for intelligence agencies of a non-American country in the top ten by GDP. They reported it. A malicious customer on that shared cluster could have seen what those clients were trai
-
Will AI Replace Knowledge Work in the Next Few Years?
Dylan Patel says AI will not replace knowledge work in the next few years. He splits a far-off 2035 picture, which he calls Looney Tunes, from 2028 and the years just ahead. In that nearer stretch, AI can be very powerful and still not better than humans in most ways. The highest-agency people then do the most interesting work and make much larger impacts on the world.
If AI is better than everyone, that picture changes. Knowledge work is in serious trouble, and what is left is what people value in each other: playing, acting, and the things humans appreciate only from other humans. He sets that later case aside from the years just ahead.
-
Why Do AI Labs Keep Their Best Models Internal?
Dylan Patel says frontier labs are keeping months of new-model progress inside the company, because once a model is public other people can distill it. Anthropic had Mythos ready in February. What the public got, many months later, is a limited version called Fable. The full Mythos model is still not out. Safety cards noted two other models after Mythos.
OpenAI has held finished models back the same way. One was done a couple of months before release. Another was shown solving math problems in a blog post and still was not released. The public gap versus Chinese open-source models is still about six months, and the labs have kept further months internal. Anthropic is likely to release a new
-
Why Did Oracle's New Mexico AI Data Center Slip?
Dylan Patel says Oracle's New Mexico data center slipped because the pipelines were not going to arrive, and the site cannot be built without them. SemiAnalysis told clients in April that the delay was coming, then again in May and June. Oracle later announced the setback and the stock fell. OpenAI has contracted a huge data center with Oracle there, and the delay could run multiple years. A Wisconsin site has slipped as well.
ClusterMax is not a stock call. It scores how reliable, secure, and usable a rented GPU cluster is, and on that score Oracle is still one of the top neoclouds. Patel calls the company solvent. The OpenAI deals were signed at a fixed margin, so Oracle cannot collect th
-
How Much Is Nvidia Backstopping Neoclouds?
Jordan Nanos says SemiAnalysis is tracking $588 billion of off-balance-sheet backstops that Nvidia is supporting. The figure is in ClusterMax and in an article the firm calls Nvidia's Backstop Universe. Those backstops help neoclouds raise the money to deploy GPUs that customers are already renting.
Nvidia does not want its revenue to depend on the hyperscalers, and it does not want customers stuck buying chips from only three or four of them. It signs on as an investment-grade offtaker so lenders can see a return. When the time comes for Nvidia to take those GPUs for its own research, another customer often wants them at a higher price. The GPUs go to that customer, and Nvidia moves on to
-
How Much Revenue Does $1 Trillion of AI Data Center Spending Need?
Dylan Patel says a trillion dollars spent today on an AI data center and GPUs needs about $1.5 trillion of revenue over six years if payback is a straight line, about $250 billion a year. The six-year clock is the accounting life of the assets. The buildings may last longer than the GPUs. Early years can bring in less and later years more, as long as the total pays the project back.
Patel calls published estimates of the required revenue too low for the pace of the build-out, and he expects revenue to come in higher than those forecasts by 2031. Alex Kantrowitz cited a Bain estimate that hyperscalers need $6 trillion a year by 2031 to break even. Meta can stop buying new capacity and still
-
Can AI Infrastructure Spending Keep Growing Exponentially?
Dylan Patel says AI capital spending cannot keep compounding at triple-digit rates from here. The build-out runs out of things to organize. The biggest neoclouds and hyperscalers already have trouble marshaling the resources for another gigawatt on top of what they are building, and raising the capital is hard even where the returns already look strong. The amounts are in the trillions of dollars.
The revenue math forces the same slowdown. Anthropic is going from under $10 billion of revenue to north of $100 billion in a single year, a 10x jump, and that pace does not produce a trillion dollars of revenue by the end of next year. Nvidia's revenue is still about doubling year on year, not tr
-
Which AI Lab Has the Least Compute?
Dylan Patel says DeepMind has the least compute of Anthropic, OpenAI, and DeepMind by the end of the year. Google is producing significantly more compute year on year, but the amount going to DeepMind is not rising as fast. The percentage allocated to model training and deployment has fallen over time, and the last six months received less than two or three years ago.
Google's business model is evolving so the company can make money at every layer of the stack. Google is physically selling a lot of chips, including TPUs, to Anthropic rather than only renting them out, and helping Anthropic on that scale hurts DeepMind. Gemini needs a niche that drives token sales, because nobody wants the t
-
Why Do So Many Neoclouds Fail ClusterMax?
Jordan Nanos says most neoclouds still miss the ClusterMax bar because a seller's market rewards raising money, building data centers, and deploying chips faster than improving the service. Almost two years after the first ClusterMax, and after Semi Analysis coined the term neocloud, many providers have not moved toward what their customers already ask for.
Semi Analysis tracks 323 providers in the Cloud TCO model. Only some are large enough to matter, and fewer pass a recommendation bar that is quite low. The two top failures are security and reliability, and both are dead simple. Companies can still run great businesses without a service customers enjoy and recommend to other people.
Onl
-
Could Anthropic Be Worth $10 Trillion?
Dylan Patel says Anthropic could be the first company in the world with a $10 trillion valuation, and the bullish case also reaches $20 trillion. Anthropic, in that view, will be a very successful company.
A $10 trillion valuation, and especially $20 trillion, requires the multiple on revenue to fall a long way. Anthropic used to raise at 30 times revenue. The IPO sketch is about $100 billion of revenue and maybe 20 times that. At $10 trillion, revenue would need to be around $1 trillion, which could happen in 2028 or 2029. The stock market can also push the multiple around.
The old 30 times private raise does not survive at that size. Revenue has to be enormous, and investors have to pay
-
Can Hyperscalers Cancel Neocloud Contracts Each Month?
Jordan Nanos says hyperscalers and large frontier labs signing neocloud contracts generally do not have a right to cancel every month and walk away. They sign five-year and six-year terms, take the full capacity, and agree to a set hourly rate and a total contract value.
Neocloud providers find the terms attractive because they have already secured land and power, built a site, and in some cases purchased the chips before the contract. The potential to slow down, pause, or cancel that Dylan Patel described applies more to self-build sites. Meta has massive campuses meant as a base load of capacity going into the future, and Meta can buy chips for them or leave an empty shell for years.
Con
-
Is Anthropic Profitable After Compute Costs?
Dylan Patel says Anthropic is now profitable once revenue is set against compute cost, with training and inference combined. OpenAI gets there within the next couple of quarters, maybe two or three, on the tokenomics model. Those companies are already breaking even.
Amazon, which he calls the biggest builder of AI infrastructure, already has profitable AI infrastructure investments. OpenAI, Anthropic, and the other customers could go insolvent, but Amazon is solvent on those contracts and making very strong gross margins. In the most recent quarter, AI revenue from Bedrock pushed Amazon's gross margin up.
AI revenue still has to grow a lot. Automating the work already in reach, and buildin
-
How Big Is Next Year's AI Capital Spending?
Dylan Patel says next year's capital spending for the AI build-out in the United States is on the order of $2 trillion. The figure counts data centers and chips and also the rest of the supply chain people are investing in. Gross domestic product is close to $30 trillion, so the spending comes to about 5 or 6 percent, not 3 or 2 percent.
A chart comparing the build-out with railroads and highways put the median year at about 3.6 percent of gross domestic product, versus 2.2 percent for the railroads, which would already make the AI build-out about 50 percent larger as a share of the economy. The published figures are still lower than the actual pace, because they understate capital spending